Predictive surrogates could cut quantum computing measurement overhead by more than 99.97%

Quantum computers, systems that process information leveraging quantum mechanical effects, have the potential of outperforming classical computers on some tasks. Despite their potential, the use of these systems remains very limited, due to their high cost and other challenges that have so far prevented their large-scale fabrication.

Researchers at the Henan Key Laboratory of Quantum Information and Cryptography and Nanyang Technological University have developed predictive surrogates, new computational models that can learn and reproduce the outputs of quantum processors.

These models, introduced in a paper published in Nature Communications, could be used to extract useful information from quantum computers and perform computations more efficiently with provable guarantees, even if users do not have direct access to advanced and expensive quantum computing hardware.

"Quantum processors have improved quickly in recent years, raising hopes that they could help solve important problems in fundamental science, chemistry, materials science, and beyond," He-Liang Huang, senior author of the paper, told Phys.org. "Yet two major obstacles still stand in the way of their practical use. The first is access. The second obstacle is speed."

A key limitation of existing quantum computers is that fabricating them and ensuring that they operate reliably over time is extremely expensive. For this reason, only a few of these computers exist worldwide. As a result, very few researchers can test quantum algorithms they developed directly on quantum processors.

In addition to being rare and expensive, most existing quantum processors still process data slowly. This is partly because reliably running quantum simulation algorithms and other quantum algorithms requires the collection of numerous repeated measurements, as well as accurate circuit evaluations.

"A less visible but equally important issue is that quantum processors themselves do not run especially fast," said Huang. "In superconducting systems, for example, a full quantum circuit is often repeated only at kilohertz rates. When a task requires millions of repetitions, this quickly turns into a serious practical bottleneck."

New classical learning models that reproduce quantum processor behavior

Huang and his colleagues wanted to address some of the well-documented limitations of quantum computers. To do this, they tried to develop classical machine learning models that would learn the typical behavior of a specific quantum processor, so well that it could accurately predict its outcomes.

"We developed a new framework called predictive surrogates," explained Huang. "A key feature of this framework is that it comes with rigorous theoretical guarantees. Once trained on a relatively small amount of data collected from a quantum processor, predictive surrogates can predict the outcomes of many future computations entirely on the classical side.

"In this way, they offer a practical route to quantum utility by cutting repeated hardware queries and measurement costs, while making advanced quantum processors useful to a much broader community."

Predictive surrogates, the algorithms developed by this research team, can essentially be seen as a quantum processor's "digital twins." By analyzing a small dataset collected from a quantum processor, these algorithms learn the relationship between the classical inputs fed to the processor and its corresponding measurements or outputs.

"Once trained, the surrogates can predict the results of many new quantum computations entirely on the classical side," said Huang. "The main advantage is efficiency. Carrying out utility-scale quantum applications on a quantum processor often requires substantial experimental time and measurement resources. Predictive surrogates can greatly reduce this burden by replacing many costly hardware evaluations with fast classical predictions."

A further advantage of predictive surrogates is that they are not black-box machine learning models. This essentially means that the processes underlying their predictions are not unknown, as the team rigorously delineated factors contributing to prediction errors, such as the dimension of classical inputs or noise in the system.

Importantly, the performance of predictive surrogates is insensitive to the quantum system size, underlying their potentials to handle large-scale quantum processors with thousands of qubits.

Initial assessments of predictive surrogates

To evaluate the potential of their proposed algorithms, Huang and his colleagues trained them using data collected from a real superconducting quantum processor with up to 42 programmable qubits. This is a type of quantum computing chip fabricated using superconductors (i.e., materials that exhibit an electrical resistance of zero under a specific critical temperature).

"We then applied them to two representative tasks: accelerating the optimization of variational quantum eigensolvers and identifying non-equilibrium quantum phases of matter," explained Huang. "In both cases, the surrogates achieved high prediction accuracy while substantially reducing measurement overhead by more than 99.97%."

Remarkably, the researchers also verified that the amount of quantum data required to effectively train the predictive surrogates does not increase dramatically as a processor's size increases. This is highly encouraging, as it suggests that the algorithms could accurately predict the outcomes of both smaller and large-scale processors.

"Traditionally, every new computation must be carried out directly and entirely on quantum hardware," said Huang. "Our results show that, for many practically important tasks, it is instead possible to learn a predictive model of the processor and then reuse that model across several future computations."

The initial tests performed by Huang and his colleagues suggest that predictive surrogates can reduce the costs of performing quantum computations by several orders of magnitude. On some tasks, the algorithms were found to also outperform previously introduced methods that require greater direct access to quantum hardware.

"More broadly, we believe this work opens a path towards democratizing access to advanced quantum hardware," said Huang. "As quantum processors continue to scale up, predictive surrogates could allow many more researchers, engineers, and scientists to benefit from powerful quantum devices without needing continuous access to them."

Future possibilities and next research steps

This recent study could soon allow more researchers to leverage the computing capabilities of quantum computers when tackling various problems rooted in physics, chemistry, materials science and other disciplines. In addition, the team's algorithms could be used to simulate quantum systems of various sizes or to find optimal solutions to complex problems.

"Predictive surrogates could help researchers reduce costly hardware evaluations and repeated measurements, while also making the power of advanced quantum devices available to a much broader community," said Huang.

"We view this work as an early step toward a broader ecosystem of AI for quantum science, and we plan to continue advancing this direction from both theoretical and experimental perspectives."

Huang and his colleagues plan to continue improving and assessing their algorithms. In addition, some team members will be conducting theoretical studies aimed at better understanding how predictive surrogates work.

"First, we want to understand predictive surrogates more deeply from the perspective of learning theory, including why they can reach beyond what classical simulation can do easily, and how their performance connects to ideas from quantum resource theory," added Yuxuan Du, co-author of the paper.

"Second, we want to extend the framework beyond standard qubit-based quantum computing to other important platforms, such as continuous-variable systems and fermionic systems. Third, we want to make these surrogates more powerful and more broadly useful, so they can support a wider range of practical, utility-scale quantum tasks while still retaining strong theoretical guarantees."

From an experimental standpoint, the researchers would like to explore the possibility of using similar computational techniques for realizing fault-tolerant quantum computers and large quantum networks. Their long-term goal is to make quantum computing more accessible, scalable and effective in tackling different scientific problems.

© 2026 Science X Network